mirrors/tinygrad
2
0
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-06-24 02:14:17 +00:00
Code Issues Projects Releases Packages Wiki Activity

why were these ever called buffer (#2483)

Browse source
This commit is contained in:
George Hotz 2023-11-27 21:02:07 -08:00 • committed by GitHub
commit 756b01f46f
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
11 changed files with 11 additions and 11 deletions
Download patch file Download diff file Expand all files Collapse all files

2
tinygrad/device.py
View file

@ -18,7 +18,7 @@ class _Device:
@functools.lru_cache(maxsize=None) # this class is a singleton, pylint: disable=method-cache-max-size-none
def __getitem__(self, x:str) -> Union[Interpreted, Compiled]:
x = x.split(":")[0].upper()
return [cls for cname, cls in inspect.getmembers(importlib.import_module(f'tinygrad.runtime.ops_{x.lower()}')) if (cname.lower() == x.lower() + "buffer") and x in self._buffers][0]
return [cls for cname, cls in inspect.getmembers(importlib.import_module(f'tinygrad.runtime.ops_{x.lower()}')) if (cname.lower() == x.lower() + "device") and x in self._buffers][0]
@functools.cached_property
def DEFAULT(self) -> str:
device_from_env: Optional[str] = functools.reduce(lambda val, ele: ele if getenv(ele) == 1 else val, self._buffers, None) # type: ignore

2
tinygrad/runtime/ops_clang.py
View file

@ -33,4 +33,4 @@ class ClangProgram:
if wait: return time.perf_counter()-st
renderer = functools.partial(uops_to_cstyle, CStyleLanguage(buffer_suffix=" restrict", arg_int_prefix="const int"))
ClangBuffer = Compiled(RawMallocBuffer, LinearizerOptions(supports_float4=False, has_local=False), renderer, compile_clang, ClangProgram)
ClangDevice = Compiled(RawMallocBuffer, LinearizerOptions(supports_float4=False, has_local=False), renderer, compile_clang, ClangProgram)

2
tinygrad/runtime/ops_cpu.py
View file

@ -46,4 +46,4 @@ numpy_fxn_for_op: Dict[Op, Callable] = {
TernaryOps.WHERE: np.where,
}
CPUBuffer = Interpreted(RawNumpyBuffer, numpy_fxn_for_op)
CPUDevice = Interpreted(RawNumpyBuffer, numpy_fxn_for_op)

2
tinygrad/runtime/ops_cuda.py
View file

@ -86,5 +86,5 @@ class CUDAProgram:
end.synchronize()
return start.time_till(end)*1e-3
CUDABuffer = Compiled(RawCUDABuffer, LinearizerOptions(supports_float4=False if getenv("PTX") else True, supports_float4_alu=False, global_max = [65535, 65535, 2147483647], local_max = [64, 1024, 1024]),
CUDADevice = Compiled(RawCUDABuffer, LinearizerOptions(supports_float4=False if getenv("PTX") else True, supports_float4_alu=False, global_max = [65535, 65535, 2147483647], local_max = [64, 1024, 1024]),
CUDARenderer, compile_cuda, CUDAProgram, cuda.Context.synchronize)

2
tinygrad/runtime/ops_disk.py
View file

@ -56,4 +56,4 @@ class RawDiskBuffer(RawBufferMapped):
return instance
disk_fxn_for_op: Dict[Op, Callable] = { BufferOps.MEM: lambda x: x, UnaryOps.NOOP: lambda x: x, UnaryOps.CAST: RawDiskBuffer.cast, MovementOps.AS_STRIDED: RawDiskBuffer.as_strided }
DiskBuffer = Interpreted(RawDiskBuffer, disk_fxn_for_op)
DiskDevice = Interpreted(RawDiskBuffer, disk_fxn_for_op)

2
tinygrad/runtime/ops_gpu.py
View file

@ -108,4 +108,4 @@ class CLProgram:
return None
return None
GPUBuffer = Compiled(CLBuffer, LinearizerOptions(), OpenCLRenderer, compile_gpu, CLProgram, CL.synchronize)
GPUDevice = Compiled(CLBuffer, LinearizerOptions(), OpenCLRenderer, compile_gpu, CLProgram, CL.synchronize)

2
tinygrad/runtime/ops_hip.py
View file

@ -137,4 +137,4 @@ class HIPGraph:
update_stats(f"<batched {len(self.jit_cache)}>", self.op_estimate, self.mem_estimate, var_vals, et, buf_count=len(input_rawbuffers), jit=jit, num_kernels=len(self.jit_cache))
return et
HIPBuffer = Compiled(RawHIPBuffer, LinearizerOptions(device="HIP"), HIPRenderer, compile_hip, HIPProgram, hip.hipDeviceSynchronize, graph=HIPGraph)
HIPDevice = Compiled(RawHIPBuffer, LinearizerOptions(device="HIP"), HIPRenderer, compile_hip, HIPProgram, hip.hipDeviceSynchronize, graph=HIPGraph)

2
tinygrad/runtime/ops_llvm.py
View file

@ -65,4 +65,4 @@ class LLVMProgram:
cfunc(*[x._buf if not isinstance(x, int) else x for x in bufs])
if wait: return time.perf_counter()-st
LLVMBuffer = Compiled(RawMallocBuffer, LinearizerOptions(supports_float4=False, has_local=False, has_shared=False), uops_to_llvm_ir, compile_llvm, LLVMProgram)
LLVMDevice = Compiled(RawMallocBuffer, LinearizerOptions(supports_float4=False, has_local=False, has_shared=False), uops_to_llvm_ir, compile_llvm, LLVMProgram)

2
tinygrad/runtime/ops_metal.py
View file

@ -152,4 +152,4 @@ class MetalGraph:
update_stats(f"<batched {len(self.jit_cache)}>", self.op_estimate, self.mem_estimate, var_vals, et, buf_count=len(input_rawbuffers), jit=jit, num_kernels=len(self.jit_cache))
return et
MetalBuffer = Compiled(RawMetalBuffer, LinearizerOptions(device="METAL"), MetalRenderer, compile_metal, MetalProgram, METAL.synchronize, graph=MetalGraph)
MetalDevice = Compiled(RawMetalBuffer, LinearizerOptions(device="METAL"), MetalRenderer, compile_metal, MetalProgram, METAL.synchronize, graph=MetalGraph)

2
tinygrad/runtime/ops_torch.py
View file

@ -52,4 +52,4 @@ torch_fxn_for_op: Dict[Op, Callable] = {
TernaryOps.WHERE: lambda x, y, z: torch.where(x != 0, y, z),
}
TorchBuffer = Interpreted(RawTorchBuffer, torch_fxn_for_op)
TorchDevice = Interpreted(RawTorchBuffer, torch_fxn_for_op)

2
tinygrad/runtime/ops_webgpu.py
View file

@ -42,4 +42,4 @@ class RawWebGPUBuffer(RawBufferCopyIn):
def toCPU(self) -> np.ndarray: return np.frombuffer(wgpu_device.queue.read_buffer(self._buf, 0), dtype=np.dtype(self.dtype.np, metadata={"backing": self})) # type: ignore
renderer = functools.partial(uops_to_cstyle, WGSLLanguage())
WebGpuBuffer = Compiled(RawWebGPUBuffer, LinearizerOptions(device="WEBGPU", supports_float4=False, local_max=[256, 256, 64], global_max=[65535, 65535, 65535]), renderer, lambda x: x, WebGPUProgram)
WebGpuDevice = Compiled(RawWebGPUBuffer, LinearizerOptions(device="WEBGPU", supports_float4=False, local_max=[256, 256, 64], global_max=[65535, 65535, 65535]), renderer, lambda x: x, WebGPUProgram)